Modelling Concentrations of Pharmaceuticals and Personal Care Products in a Canadian Watershed

Abstract

The concentrations of several pharmaceuticals and personal care products (PPCPs) were simulated in the Grand River watershed using the PhATE (Pharmaceutical Assessment and Transport Evaluation) model. PPCPs that were previously measured in the Grand River watershed were selected as the target compounds for this study. The application of the PhATE model was adapted to account for Canadian climatic conditions and its seasonal variability. In this regard, seasonal hydrological parameters (i.e., stream flow and velocity) were estimated based on historical data. Chemical loss parameters, including in-stream decay, human loss, and removal efficiency of treatment plants, were initially extracted from the literature and then calibrated to the site data. In general, the adapted PhATE model reasonably simulated pharmaceuticals with continuous use by humans such as ibuprofen, naproxen, carbamazepine, and gemfibrozil. Validation, which was performed only for ibuprofen, naproxen, and carbamazepine suggests that the calibrated model is able to reliably simulate concentrations. The model is the most accurate in validation when simulating carbamazepine which is the most persistent substance tested in this study, and as such, variations in its concentration primarily follow variation in stream flow.

Les concentrations de plusieurs produits pharmaceutiques et de soins personnels (PPSP) ont t simules dans le bassin de la rivire Grand l'aide du modle PhATE (Pharmaceutical Assessment and Transport Evaluation). Les PPSP qui avaient t mesurs antrieurement dans le bassin versant de la rivire Grand ont t slectionns en tant que composs cibles pour cette tude. L'application du modle PhATE a t adapte afin de tenir compte des conditions climatiques du Canada et de sa variabilit saisonnire. cet gard, les paramtres hydrologiques saisonniers (c.--d. le dbit et la vitesse dcoulement) ont t estims en fonction des donnes historiques. Les paramtres de perte chimique, y compris la dtrioration dans le cours d'eau, les pertes causes par les activits humaines et l'efficacit dpuration des stations de traitement, ont t extraits initialement de la documentation puis ont t talonns en fonction des donnes du site. En gnral, le modle PhATE adapt a permis de simuler dans une mesure raisonnable les produits pharmaceutiques avec utilisation humaine continue, comme l'ibuprofne, le naproxne, la carbamazpine et le gemfibrozil. La validation, qui n'a t excute que pour l'ibuprofne, le naproxne et la carbamazpine, porte croire que le modle talonn est capable de simuler les concentrations de manire fiable. L o le modle s'avre le plus prcis, au chapitre de la validation, c'est lorsqu'il simule la carbamazpine, substance la plus persistante teste dans la prsente tude. ce titre, les variations dans sa concentration suivent principalement la variation du dbit du cours d'eau.

Introduction

Pharmaceuticals and personal care products (PPCPs) are chemicals used extensively in society for treatment, prevention, and beautification. Humans ingest pharmaceuticals and related products almost daily, and a fraction of each dosage is excreted (i.e., as feces or urine) due to incomplete metabolism in the human body. This excreted fraction may ultimately be discharged to surface waters via waste water treatment plant discharges (Lissemore et al., 2006; Miao et al., 2002; Metcalfe et al., 2003; Kreisberg, 2007; Buser et al., 1999; Heberer, 2002; Ternes, 1998; Andreozzi et al., 2003; Bendz et al., 2005; Clara et al., 2005; Lindqvist et al., 2005; Nakada et al., 2007; Ying et al., 2009).

As a result of these discharges, detectable amounts of PPCPs have been reported in surface and ground waters and drinking water (Daughton and Ternes, 1999; Heberer, 2002; Buzby, 2005; Ternes et al., 1999; Carrara et al., 2008; Godfrey et al., 2007; Kolpin et al., 2002; Lissemore et al., 2006; Metcalfe et al., 2003; Miao et al., 2002). Canadian watersheds have been found to be impacted (Lissemore et al., 2006; Metcalfe et al., 2003; Miao et al., 2002; Brun et al., 2006; Hao et al., 2006; Lishman et al., 2006), although detected concentrations have generally been low, i.e., between nanograms and micrograms per liter (Lissemore et al., 2006; Ternes et al., 1999; Tixier et al., 2003). However, even at low concentrations, the continual discharge of these chemicals into the environment may have adverse health effects on aquatic biota, such as feminization of various species (Barcelo and Petrovic, 2007; Jobling et al., 1998). Therefore, the fate and transport of PPCPs in varying environments has emerged as an important research topic (Barcelo and Petrovic, 2007; Kreisberg, 2007).

The detection and identification of PPCP compounds in natural systems requires highly sensitive instruments, and these analyses consume considerable time and money. Therefore, there has been an increasing interest in the development of models capable of reliably predicting the fate of pharmaceuticals. For example, the GREAT-ER (Geo-referenced Regional Exposure Assessment Tool for European Rivers) and PhATE (Pharmaceutical Assessment and Transport Evaluation) models were developed to estimate the concentrations of chemicals in surface waters (Cunningham et al., 2010; Anderson et al., 2004; Feijtel et al., 1997; Cunningham, 2008; Lindberg et al., 2006; Robinson et al., 2007). The models can be employed to facilitate an assessment of the potential toxicity effect of pharmaceuticals on aquatic life (Cunningham, 2008; Hannah et al., 2009). Furthermore, they can help guide the design of cost-effective field sampling strategies by highlighting the compounds with higher potential risk (Cunningham, 2008; Hannah et al., 2009). While these models were developed in European and United States settings, the climate in Canada exhibits considerable seasonal variability, and the long cold season may affect the fate of pharmaceuticals and personal care products in the environment. Exploring whether the use of seasonal parameters could improve PhATE predictions was the main objective of this research. Therefore, in this study, the PhATE model was adapted to account for Canadian conditions by seasonally predicting concentrations of aquatic chemicals (i.e., spring, summer, fall, and winter). The Grand River watershed, located in Ontario, Canada, was used as a case study to evaluate the performance of the adapted PhATE model.

Methods

In this study, the PhATE model was employed in a manner that allowed various physical, chemical, and hydrological factors to vary with the seasons (i.e., spring, summer, fall, and winter). Initially, compound-specific values for model parameters, such as removal efficiency of treatment plants, in-stream loss, and human loss, were collected from the literature. After an assessment of the model performance, the parameters were calibrated using available data that were collected from the Grand River watershed. Observation data not used for calibration served as the basis for subsequent model validation, to verify whether the calibrated model could adequately predict a separate data set of PPCP concentrations.

PhATE Model

The PhATE (Pharmaceutical Assessment and Transport Evaluation) model was developed to simulate concentrations of active pharmaceutical ingredients in eleven watersheds across the United States (Anderson et al., 2004). The PhATE model is based upon mass balance equations that describe contaminant fate in each segment of the watershed. For each segment, mass enters either via inflow from upstream segments or from wastewater treatment plants as point sources along the segment. Contaminants leave the segment via first-order in-stream decay, flow diversions, or outflow to a downstream segment (Anderson et al., 2004). The mass of PPCPs entering from a given wastewater treatment plant is estimated as the product of the average annual human usage of the compounds per capita multiplied by the size of the population served by the wastewater treatment plants. This loading is reduced via two loss terms, namely: (1) removal by human metabolism; and (2) removal within the treatment plants (Anderson et al., 2004).

Grand River Watershed

The Grand River watershed is located in southwestern Ontario and drains an area of about 6,800 km² (Hao et al., 2006). The watershed receives treated effluent from a total of forty wastewater treatment plants which serve about 530,000 residents ( Figure 1 ). Twenty eight are municipal plants and, of these, fifteen provide advanced tertiary treatment, nine provide conventional secondary treatment, and four are lagoons (Koycheva, 2003).

Figure 1. Locations of wastewater treatment plants in the Grand River Watershed.

Target Chemicals

The selection of compounds in this study was limited to those where seasonal measured data were available to allow a robust assessment of the model. These consisted of ibuprofen, naproxen, carbamazepine, gemfibrozil, sulfamethoxazole, trimethoprim, and bezafibrate. These compounds are the most frequently detected compounds in Canadian surface waters (Lissemore et al., 2006; Metcalfe et al., 2003; Brun et al., 2006; Hao et al., 2006; Lishman et al., 2006). The compounds span a range of persistence in wastewater treatment with naproxen and ibuprofen reported to have relatively high removals ranging from 89% to 99% (Metcalfe et al., 2003; Lishman et al., 2006), while carbamazepine has been reported to be recalcitrant (Metcalfe et al., 2003; Brun et al., 2006).

A thorough review of the literature was conducted to obtain model parameters for the target compounds ( Table 1 ). From Table 1 it can be seen that in almost all instances there is a wide range of removals reported for the target compounds. However, in general, when the removals are compared between compounds, the patterns in removal relative to each other are consistent across processes. The wide range of parameter values for compounds presents considerable challenges, and hence, necessitates the staged model calibration that is described in this paper. As will be subsequently described, the literature values in Table 1 were used to conduct an initial round of simulations and to constrain the auto-calibration.

Table 1. Literature-derived ranges of parameter values for the selected compounds.

Compounds	Lagoon R^1	References	Sec. TP^2	References	Ter. TP^3	References	In-Stream Loss	References	Human Loss	References
Ibuprofen	77100	(Conkle et al., 2008; Metcalfe et al., 2003; Ying et al., 2009)	7297	(Castiglioni et al., 2006; Jones et al., 2007; Kasprzyk-Hordern et al., 2010; Lee et al., 2003; Paxeus, 2004; Santos et al., 2007; Ternes, 1998; Yu et al., 2009)	81100	(Bendz et al., 2005; Nakada et al., 2007; Paxeus, 2004; Roberts and Thomas, 2006; Ying et al., 2009; Yu et al., 2006, 2009)	0.0221.124	(Fono et al., 2006; Kunkel and Radke, 2008; Matamoros et al., 2009; Tixier et al., 2003; Yamamoto et al., 2009; Yu et al., 2006, 2009)	6192	(Lienert et al., 2007; Ternes, 1998)
Naproxen	9099.8	(Conkle et al., 2008; Metcalfe et al., 2003; Ying et al., 2009)	32.299	(Kasprzyk-Hordern et al., 2010; Lee et al., 2003; Metcalfe et al., 2003; Paxeus, 2004; Santos et al., 2007; Ternes, 1998)	5099.8	(Bendz et al., 2005; Joss, 2005; Lindqvist et al., 2005; Metcalfe et al., 2003; Nakada et al., 2007; Paxeus, 2004; Ying et al., 2009; Yu et al., 2006)	0.05111.885	(Fono et al., 2006; Kunkel and Radke, 2008; Leifer, 1988; Packer et al., 2003; Radke et al., 2010; Tixier et al., 2003; Yu et al., 2006)	8999	(Lienert et al., 2007; Runkel et al., 1973; National Library of Medicine (NLM), 2010)
Carbamazepine	551	(Conkle et al., 2008; Metcalfe et al., 2003; Ying et al., 2009)	053	(Castiglioni et al., 2006; Kasprzyk-Hordern et al., 2010; Paxeus, 2004; Santos et al., 2007; Ternes, 1998; Yu et al., 2009; Zhang et al., 2008)	060	(Bendz et al., 2005; Nakada et al., 2007; Paxeus, 2004; Ying et al., 2009)	0.00120.24	(Andreozzi et al., 2003; Matamoros et al., 2009; Tixier et al., 2003; Yamamoto et al., 2009; Zhang et al., 2008)	6987	(Lienert et al., 2007; Zhang et al., 2008; NLM, 2010)
Gemfibrozil	1560	(Conkle et al., 2008; Metcalfe et al., 2003; Ying et al., 2009)	581	(Lee et al., 2003; Paxeus, 2004; Ternes, 1998; Yu et al., 2009)	5575	(Bendz et al., 2005; Paxeus, 2004; Ying et al., 2009; Yu et al., 2006)	0.05781.124	(Fono et al., 2006; Lin et al., 2009; Yu et al., 2006)	7494	(Lienert et al., 2007; Smith, 1976)
Sulfamethoxazole	1799	(Conkle et al., 2008; Yu et al., 2009)	999.8	(Castiglioni et al., 2006; Gobel et al., 2007; Karthikeyan and Meyer, 2006; Kasprzyk-Hordern et al., 2010; Yu et al., 2009)	199	(Bendz et al., 2005; Gobel et al., 2007; Nakada et al., 2007)	0.05911.09	(Andreozzi et al., 2003; Lam and Mabury, 2005; Trovo et al., 2009; Werner et al., 2005)	8090	(Gobel et al., 2005; Lienert et al., 2007)
Trimethoprim	6599	(Yu et al., 2009)	0100	(Brown et al., 2006; Choi et al., 2008; Gobel et al., 2007; Gulkowska et al., 2008; Kasprzyk-Hordern et al., 2010; Lindberg et al., 2006; Paxeus, 2004; Watkinson et al., 2007; Yu et al., 2009)	099.6	(Bendz et al., 2005; Gobel et al., 2007; Lindberg et al., 2006; Nakada et al., 2007; Paxeus, 2004; Roberts and Thomas, 2006)	0.03960.24	(Batt et al., 2006; Werner et al., 2005)	5056	(Gobel et al., 2005)
Bezafibrate	0100*		0100	(Castiglioni et al., 2006; Kasprzyk-Hordern et al., 2010; Lindberg et al., 2006)	1593	(Lindqvist et al., 2005)	00.161	(Kunkel and Radke, 2008; Radke et al., 2010)	4950	(Lienert et al., 2007; Ternes, 1998)
^1 Removal efficiency by lagoon.
^2 Secondary treatment plant removal.
^3 Tertiary treatment plant removal.

Hydrological Properties

The PhATE model employs information on average and low flows and the corresponding velocities for each stream segment. Seasonal hydrological parameters were estimated using flow values reported by the Grand River Conservation Authority (GRCA) and the Water Survey of Canada (WSC). In this study, the hydrological season was defined as three continuous months of relatively similar flows, and hence, they were categorized as follows: winter December, January, and February; spring March, April, and May; summer June, July, and August; and fall September, October, and November.

The seasonal mean and low flow for each stream segment was estimated using a relationship between flow and drainage area. Drainage area, precipitation, and temperature are the most important variables for estimating average annual flow (Vogel et al., 1999), and of these variables, drainage area is recognized as the most influential factor (Mohamoud and Pamar, 2006). In the Grand River watershed, drainage area is highly correlated with average and low seasonal flows and the regressions yielded R² values ranging from 0.78 to 0.99 and 0.69 to 0.99, respectively.

The velocity in each segment was calculated using available flow data paired with relevant morphologic features of the stream segments reported in previous studies (Annable, 1996). The velocities were estimated using a logarithmic relationship between velocity and flow rate (Leopold and Maddock, 1953). Low-flow velocities for each stream segment were calculated using Equation 1 (United States Environmental Protection Agency (USEPA), 1997).

where V_low is low flow velocity, V_avg is mean flow velocity, Q_low is low flow, and Q_avg is mean flow.

Numerical Experiments

As previously described, an initial survey of the literature was performed in order to establish a reasonable range of expected values for these parameters. PhATE simulations were then conducted using average values for the parameters that were based on the literature ranges. In this part of the study, seasonal hydrologic parameters were incorporated into four input files, with one input file per season. Also included in each input file were corresponding seasonal estimates of pharmaceutical loading and the various loss parameters. The values of the parameters for PPCPs and endocrine-disrupting chemicals (EDCs) that were included in the individual files were categorized seasonally according to the reported time of the measurement, if it was available, and also involved the application of some expert judgment.

As will be subsequently demonstrated, the simulations with the literature-derived parameter values yielded unrealistic predictions for certain compounds (e.g., naproxen) when applied to the Grand River watershed. Hence, in the second stage of the study an automated calibration was conducted to obtain parameter values that best match observed behaviour in the Grand River watershed. Automated calibration was used to adjust uncertain model parameters in order to obtain the best possible correspondence between model outputs and historical observation data. Parameters including flow and velocity of stream segments and annual per capita pharmaceutical usage were derived from data that were readily available for the Grand River watershed and/or Canada, and hence, were not adjusted. Conversely, the removal efficiencies of lagoons, secondary and tertiary treatment plants, in-stream loss, and human loss were not known with certainty, and hence, were varied during the calibration exercise. A total of 94 data points were available for a location of the Grand River that is near Mannheim; these had been collected over a period that spanned from November 2006 to August 2008 (Halle, 2009). These concentrations of PPCPs served as the observational data set and were utilized for model calibration. Note that PhATE assumes steady-state, and thus, the model had to be re-run for each data point in the flow and concentration time-series.

In this project, the DDS (Dynamically Dimensioned Search) optimization algorithm (Tolson and Shoemaker, 2007) through the OSTRICH (Optimization Software Tool for Research In Computational Heuristics) interface (Matott, 2005) was employed. OSTRICH is specifically designed so that it can easily be linked with models that accept and generate text-based input/output files with no user interaction required. However, the PhATE model works only through a graphical user interface (GUI) and depends heavily on user-interaction via various button clicks and dialog windows. Furthermore, the PhATE model utilizes the Access Database file format for its input/output files. Therefore, the OSTRICH software was modified to support the Access Database file format in order to link OSTRICH with the PhATE model. Also, to programmatically manipulate the input of the PhATE model, AutoHotKey (http://www.autohotkey.com, an open-source utility for windows that can simulate keystrokes and mouse clicks) was used to allow OSTRICH to run the PhATE model without user intervention.

Calibration was conducted using the average parameter values as the initial solution and was constrained by the parameter values reported in the literature. Calibration required definition of an objective function, and in this study minimizing the conventional weighted sum of squared deviations of simulated concentration values from measured data was utilized (Hill and Tiedeman, 2007). Equal importance (weights) was assumed for all observations while calibrating for all compounds except for bezafibrate and sulfamethoxazole. For calibration of these two compounds, the measured concentrations above the detection limit received a greater weight (i.e., a weight twice as large as the weight of observations below the detection limit).

Two hundred function evaluations were employed in the auto-calibration process. Each function evaluation consisted of n PhATE model runs, where n is the number of available observations in each season. For example, for the ibuprofen calibration in winter, each function evaluation consisted of 20 PhATE model runs, one for each of the input files. For this study, a computational budget of 200 function evaluations (or about 4.5 hours) was assigned to each calibration trial, with multiple trials required for calibrating seasonal behavior.

Following calibration, a validation exercise assessed the ability of the calibrated model to adequately predict PPCP concentrations. A limited number of seasonal data was available for validation, and hence, the validation was performed for three compounds, ibuprofen, naproxen, and carbamazepine. The validation data included the concentrations that were collected at two different locations in the Grand River watershed, near Mannheim and Holmedale (Kormos, 2007). Two samples had been collected at each location monthly from April 2005 to March 2006 (Kormos, 2007). These pharmaceuticals were simulated using the seasonal calibrated parameters and actual flows at the time of sampling. The seasonal simulated concentrations were compared to the seasonal measured data and were deemed to have a favorable match if the predicted values were within a factor of ten of the measured data (Anderson et al., 2004). The value of the sum of squared deviations of simulated concentration values from measured data was also employed to assess the goodness of fit of the model during validation.

Results

Simulations with Literature-Derived Parameter Values

The cumulative distribution of the measured concentrations was used to assess the ability of the model to reliably predict the range of concentrations that were observed for each season. As an example, Figure 2 shows the predicted concentrations (PECs) of ibuprofen in the spring and naproxen in the fall along with the cumulative distribution function of the corresponding measured concentration values. The results showed that for ibuprofen and naproxen about 50% of the observed values in each season fall within the range of the simulated concentrations associated with average seasonal mean flow and average seasonal low flow values. However, in the fall, only 2030% of ibuprofen and naproxen measured concentrations fall within the corresponding simulated concentration ranges. The simulated concentrations of carbamazepine show good agreement with the measured data. Only one measured data point in the spring and one in the summer exceed the low flow-based simulated concentration.

Figure 2. Observed versus predicted concentrations using literature-derived parameters PEC: predicted concentration, a) ibuprofen in spring, b) naproxen in fall.

Auto-Calibration Results

The five PhATE model parameters (i.e., removal efficiencies in lagoon, secondary and tertiary treatment plants, in-stream decay, and human loss) were seasonally calibrated to obtain model results that best match the observed concentrations in the Grand River watershed. The PhATE model performance before and after auto-calibration was compared for each compound. As examples of this comparison, the simulated concentrations of ibuprofen and naproxen are shown versus the corresponding measured data in Figure 3 . From these graphs it can be observed that the PhATE model is capable of predicting the concentrations of the PPCPs when actual flows are used, and calibration results in a significantly better estimation (compared to pre-calibration) of the concentrations in the Grand River watershed. The PhATE model is able to better simulate ibuprofen, naproxen, carbamazepine, and gemfibrozilas compared to other compounds. The estimated R² values for these drugs ranges from 0.48 for ibuprofen to 0.69 for gemfibrozil.

Figure 3. Observed and simulated concentrations before and after auto-calibration, a) ibuprofen, b) naproxen.

Table 2 shows the calibrated parameter values, which were referred to as the most probable seasonal parameters of the selected compounds for the case study. As shown in Table 2 , for the majority of pharmaceuticals, the removal efficiency of the tertiary treatment plants has the highest removal rates among the removal efficiencies of the different treatment plants. However, for ibuprofen and naproxen, lagoons are found to be the most efficient treatment plants. The efficiency of the treatment plants in removing naproxen, carbamazepine and bezafibrate varies seasonally. For example, naproxen has the highest removal efficiency in the treatment plants in the summer season. However, there is no significant seasonal variability in the removal efficiencies of the treatment plants for ibuprofen and trimethoprim. Table 2 also shows that the calibrated human loss values for each compound varies modestly between seasons. It might be expected that human loss would be essentially constant for all seasons. Hence, the variation is attributed to uncertainties in modelling and calibration and/or seasonality in loading. As loading was fixed in the calibration step for each season, seasonality in loading would be reflected in variability in the human loss parameter.

Table 2. Calibrated parameter values.

		Parameters
Compound	Season	Removal by Lagoon (%)	Removal by Secondary TP (%)	Removal by Tertiary TP (%)	In-Stream Loss (day^1)	Human Loss (%)
Ibuprofen	Winter	78	84	87	0.083	89
	Spring	97	86	84	1.093	76
	Summer	79	93	82	0.418	83
	Fall	98	93	87	0.002	79
Naproxen	Winter	95	32	52	2.150	66
	Spring	91	32	52	2.314	66
	Summer	90	42	78	1.154	64
	Fall	98	33	55	0.663	80
Carbamazepine	Winter	49	37	39	0.237	82
	Spring	0	12	68	0.238	83
	Summer	33	21	24	0.005	79
	Fall	13	21	45	0.238	85
	Spring	0	95	82	1.437	47
	Summer	44	81	92	1.020	0
	Fall	32	89	86	0.120	57
Gemfibrozil	Winter	15	45	62	0.290	75
	Spring	60	33	55	0.339	75
	Summer	60	18	69	1.095	87
	Fall	15	41	75	0.288	87
Bezafibrate	Winter	2	0	16	0.003	40
	Spring	0	1	15	0.046	40
	Summer	50	68	68	0.161	45
	Fall	33	33	68	0.141	40
Sulfamethoxazole	Winter	No calibration, as all the measured data were below the detection limit
	Spring	17	18	22	0.861	80
	Summer	21	23	86	0.201	86
	Fall	17	20	60	0.683	80
Trimethoprim	Winter	65	42	92	0.04	53
	Spring	65	65	92	0.039	50
	Summer	66	84	93	0.038	53
	Fall	65	79	93	0.04	56

For some cases, the calibrated parameter values seem to be unreasonable; for instance, high variability in in-stream decay value during a year (e.g., two or three orders of magnitude), a large difference between removal efficiencies of a treatment plant between seasons, and lower removal rates in summer than in winter. Potential explanations for these patterns include: (1) the calibration was performed only at one sample location which was mostly affected by secondary treatment plant removal; therefore, in-stream decay, lagoon, and tertiary treatment plant removal were not sufficiently influential parameters for the calibration; (2) a high percent of observed concentrations for some of the compounds were reported below detection limit (e.g., bezafibrate); and (3) correlation between parameters in the calibration exercise.

In order to assess the significance of applying seasonality to the model parameters in this study, two scenarios, annual and seasonal, were considered. In the annual scenario, all the hydrological and chemical parameters were assumed to be constant over the year and annual mean flows were employed. In the seasonal scenarios the hydrological and chemical parameters were assumed to vary with season. The average annual and seasonal simulated concentrations of ibuprofen and naproxen are plotted in Figure 4 along with the measured concentrations. The comparison is based on the average simulated and measured concentrations for each season. As shown in Figure 4 , incorporating seasonality of parameters improves the simulation of ibuprofen and naproxenwhen compared to the observed data.

Figure 4. Comparison of annual versus seasonal simulations, a) ibuprofen, b) naproxen.

The improvement in seasonally predicting pharmaceutical concentrations was further assessed by comparing the percentage of the field measured concentrations that were between the simulated mean and low flow concentrations for each season. The results indicate that simulation of ibuprofen and naproxen improves after seasonal calibration in most of the seasons, as a larger percentage of the measured data is between the simulated seasonal mean and low flow concentrations. In the winter, the percent of measured data between seasonal mean and low flow concentrations increases from 55 to 60% for ibuprofen and from 4 to 50% for naproxen. While the use of seasonal parameters for naproxen improves simulations in the winter and the summer, the number of measured concentrations within the two predicted concentrations in the spring and the fall decreases. A higher number of measured concentrations of carbamazepine fall between the simulated concentrations when annual parameters are applied.

Beside seasonality of parameters, other factors may improve the simulation results. There are good fits between measured and calibrated concentrations when the flow rates that were recorded at the time of sample collection are used, although the average simulated concentrations are less than the average measured concentrations for all of the seasons and compounds. This result suggests that the average flow might not be the best strategy in estimating the average concentrations, and as lower flow values would typically result in higher concentrations, other types of central tendency measures, like the median flows, which are typically less than average streamflows, might be preferred. Figure 5 compares the average and median flow values at the time of sampling with the average estimated flows, and shows the median flow rates were considerably lower than the average flow. As the estimated concentrations in all cases are less than the average measured concentrations, accounting for the median flow instead of mean flow is expected to increase the simulated concentration values and improve the seasonal simulations.

Figure 5. Comparison of average and median flows versus season.

Validation Results

In the validation phase, ibuprofen, naproxen, and carbamazepine were simulated using the seasonally calibrated parameters and actual flows at the time of sampling. Figure 6 shows the simulated concentrations of ibuprofen, naproxen, and carbamazepine along with the corresponding measured concentration values. For carbamazepine, as shown in Figure 6 , there is a good fit between the simulated and the measured concentrations at both sampled locations, with the exception of the August sample event at Holmedale, when the measured concentration was extremely high (likely an outlier), about 1 g/L. For ibuprofen, the simulated concentrations and the measured concentrations at Mannheim have a good fit, while the simulations at Holmedale do not match the measured data sufficiently well. The better model performance for Mannheim is probably due to the fact that the calibration was also performed on this segment, while Holmedale was not considered in calibration. Note that this behavior (different accuracies in the Grand River near Mannheim and Holmedale) is not observed for carbamazepine, possibly because this compound is relatively persistent in the environment. Therefore, the variation in carbamazepine concentration is generally governed by the variation in the flow rate. The validation results for naproxen are not as good as the validation results for ibuprofen and carbamazepine at both locations. The sum of squared deviations (SSD) of simulated concentration of naproxen from measured data at Mannheim was 2,521 which is one order of magnitude larger that ibuprofen SSD (485) and carbamazepine SSD (332).

Figure 6. Observed and simulated concentrations model validation.

Although the accuracy in validation might not seem very high, previous researchers have suggested that simulations of this class of chemicals are reasonable if the predicted data are within a factor of ten of the measured data (Anderson et al., 2004). While the degree of model accuracy that is deemed to be acceptable will depend upon the intended use of the simulations, the validation results in this study indicate relatively good fits between observed data and corresponding model output.

Conclusions

This study employed the PhATE model to predict transport of frequently detected pharmaceuticals, personal care products, and endocrine disruptors in the Grand River watershed. When the application of the PhATE model was adapted to reflect seasonal conditions it was found to be capable of accurately simulating pharmaceutical concentrations in this watershed. In the pre-calibration phase (i.e., using previously published parameter values), the model tended to over predict concentrations.

After calibration, the PhATE model accurately predicted pharmaceutical concentrations in the Grand River. The use of accurate values for stream flow and velocity was found to be important for estimation of PPCPs in the surface water. In the validation phase, the performance of the calibrated PhATE model was tested with a set of data that was not used in calibration at two sampling sites. The validation results showed relatively good fit between the measured data and corresponding model output, and the deviations of the simulated concentrations from the measured data were consistently less than a factor of ten.

Acknowledgements

The authors acknowledge Webber Chan who modified the OSTRICH software to support Access Database and the use of AutoHotkey.exe to run the PhATE model automatically.